Predicting Wildfires¶
By: Eugenia Poon, Kayla Zhu
Motivation: The progressing issue of California wildfires in recent years has destroyed ecosystems, magnified air pollution and health risks, and weakened labor and resources.
Goal: predict the risk/presence/spread of fire
Table of Contents¶
Load and Clean Data
Feature Engineering
Modeling
In [1]:
%matplotlib inline
import os
import os.path
import random
import warnings
import zipfile
import geopandas as gpd
import numpy as np
import pandas as pd
import seaborn as sns
import tensorflow as tf
import xarray as xr
from IPython.display import HTML, Javascript, display
from imblearn.over_sampling import SMOTE
from imblearn.pipeline import Pipeline
from imblearn.under_sampling import RandomUnderSampler
from matplotlib import pyplot as plt
from plotly import express as px, graph_objects as go
from sklearn.metrics import (
balanced_accuracy_score,
f1_score,
mean_absolute_error,
mean_squared_error,
precision_recall_fscore_support,
)
from sklearn.preprocessing import StandardScaler
from tensorflow.keras.layers import Dense, Input, Normalization
from tensorflow.keras.models import Model
from tensorflow.keras.utils import to_categorical
from xgboost import XGBClassifier
from wildfireFunc import (
getDF,
getFireRange,
getInter,
getNC,
plotInter,
plotRes,
model,
plot_loss,
score,
split,
)
warnings.filterwarnings('ignore')
plt.rcParams['figure.figsize'] = (6, 4)
plt.rcParams['figure.dpi'] = 400
plt.rcParams['font.size'] = 5
plt.rcParams['figure.titlesize'] = 10
plt.rcParams['axes.linewidth'] = 0.1
plt.rcParams['patch.linewidth'] = 0
display(HTML("<style>.container { width:100% !important; }</style>"))
random.seed(100)
tf.random.set_seed(100)
np.random.seed(100)
Load and Clean Data¶
Shapefiles¶
- Only Northern Sierra Nevada: Tahoe, Plumas, and Lassen National Forest
- How to Query
| Sierra Nevada | Fire Perimeter | |
|---|---|---|
| SOURCE | Sierra Nevada Conservancy Boundary | CAL Fire |
| OPEN | View Data Source | View Data Source |
| OPEN | California Fire Perimeters (1950+) | |
| OPEN | Query | Query |
| Where: | 1=1 | YEAR_=2021 |
| Geometry Type: | Polygon | Polygon |
| Spatial Relationship: | Intersects | Intersects |
| Units: | Meters | Meters |
| Out fields: | * | * |
| Return Geometry: | True | True |
| Format: | GeoJSON | GeoJSON |
| Units: | Meters | Meters |
| CLICK | Query (GET) | Query (GET) |
In [2]:
# sierra nevada
url = 'https://services1.arcgis.com/sTaVXkn06Nqew9yU/arcgis/rest/services/Sierra_Nevada_Conservancy_Boundary/FeatureServer/0/query?where=1%3D1&objectIds=&time=&geometry=&geometryType=esriGeometryPolygon&inSR=&spatialRel=esriSpatialRelIntersects&resultType=none&distance=&units=esriSRUnit_Meter&relationParam=&returnGeodetic=false&outFields=*&returnGeometry=true&returnCentroid=false&featureEncoding=esriDefault&multipatchOption=xyFootprint&maxAllowableOffset=&geometryPrecision=&outSR=&defaultSR=&datumTransformation=&applyVCSProjection=false&returnIdsOnly=false&returnUniqueIdsOnly=false&returnCountOnly=false&returnExtentOnly=false&returnQueryGeometry=false&returnDistinctValues=false&cacheHint=false&orderByFields=&groupByFieldsForStatistics=&outStatistics=&having=&resultOffset=&resultRecordCount=&returnZ=false&returnM=false&returnExceededLimitFeatures=true&quantizationParameters=&sqlFormat=none&f=pgeojson&token='
snBoundaries = gpd.read_file(url).to_crs(epsg=4326)
snBounds = snBoundaries.geometry.bounds
bounds = [-122, 39.35, snBounds.iloc[0].maxx, 41]
boundaries = gpd.clip(snBoundaries, bounds)
bounds = boundaries.geometry.bounds
display(snBounds)
# fire perimeter
url1 = 'https://services1.arcgis.com/jUJYIo9tSA7EHvfZ/ArcGIS/rest/services/California_Fire_Perimeters/FeatureServer/2/query?where=YEAR_%3D'
url2 = '&objectIds=&time=&geometry=&geometryType=esriGeometryPolygon&inSR=&spatialRel=esriSpatialRelIntersects&resultType=none&distance=&units=esriSRUnit_Meter&relationParam=&returnGeodetic=false&outFields=*&returnGeometry=true&returnCentroid=false&featureEncoding=esriDefault&multipatchOption=xyFootprint&maxAllowableOffset=&geometryPrecision=&outSR=&defaultSR=&datumTransformation=&applyVCSProjection=false&returnIdsOnly=false&returnUniqueIdsOnly=false&returnCountOnly=false&returnExtentOnly=false&returnQueryGeometry=false&returnDistinctValues=false&cacheHint=false&orderByFields=&groupByFieldsForStatistics=&outStatistics=&having=&resultOffset=&resultRecordCount=&returnZ=false&returnM=false&returnExceededLimitFeatures=true&quantizationParameters=&sqlFormat=none&f=pgeojson&token='
years = ['2020', '2021']
firePer = pd.concat([gpd.read_file(url1 + y + url2) for y in years])
dateCol = ['ALARM_DATE', 'CONT_DATE']
firePer = firePer[firePer.STATE == 'CA'].dropna(subset=dateCol)
firePer.geometry = firePer.geometry.make_valid() # 2021 has invalid geometry
firePer = gpd.clip(firePer, boundaries) # fires in area
firePer[dateCol] = firePer[dateCol].apply(pd.to_datetime, unit='ms')
firePer = firePer.sort_values('ALARM_DATE').reset_index(drop=True)
print(f'{len(firePer)} fires')
display(firePer.sort_values('Shape__Area')[
['FIRE_NAME', 'Shape__Area']].tail())
| minx | miny | maxx | maxy | |
|---|---|---|---|---|
| 0 | -123.039382 | 35.064862 | -117.735384 | 41.994919 |
64 fires
| FIRE_NAME | Shape__Area | |
|---|---|---|
| 26 | LOYALTON | 3.198601e+08 |
| 41 | W-5 COLD SPRINGS | 6.049077e+08 |
| 58 | SUGAR | 7.236506e+08 |
| 34 | NORTH COMPLEX | 2.183903e+09 |
| 59 | DIXIE | 6.692866e+09 |
In [3]:
fig, ax = plt.subplots(nrows=1, ncols=1, dpi=600)
snBoundaries.plot(ax=ax, color='grey', alpha=0.1, linewidth=0)
boundaries.plot(ax=ax, color='grey', alpha=0.3, linewidth=0)
firePer.plot('YEAR_', legend=True, alpha=0.6, linewidth=0, ax=ax)
plt.tick_params(axis='both', labelsize=5)
ax.set(xlabel='longitude', ylabel='latitude')
plt.show()
Geospatial Data: GRIDMET¶
- GRIDMET (Download NetCDF Data)
- Clip 2021 and 2022 GRIDMET data to boundaries within Tahoe, Plumas, and Lassen National Forest area in Sierra Nevada
Description¶
- What starts a fire? high winds, high temperatures, low humidity
- Years: 2020-2021
- Resolution: 4638.3 meters = 4.6383 km = 1/24 degrees
- Temporal Resolution: daily
- Key inputs into the NFDRS model are: fuels, weather, topography and risks.
| variable | var | units | description |
|---|---|---|---|
| relative_humidity | rmin | % | minimum relative humidity |
| air_temperature | tmmx | K | maximum temperature |
| wind_speed | vs | m/s | wind velocity at 10m |
| burning_index_g | bi | NFDRS fire danger index | burning index |
In [4]:
path = 'data/gm.nc'
if os.path.isfile(path+'.zip') == False: # ~3 minutes
getNC(path, bounds, boundaries, years)
!rm data/gm.nc
ds = xr.open_dataset(zipfile.ZipFile(path+'.zip').open('gm.nc'))
print(ds.dims)
Frozen({'lon': 52, 'lat': 44, 'day': 731})
Interpolation¶
- Used for testing data with higher factor --> less data
- To get a lower resolution, choose a higher factor
- Lowering resolution will result in misleading performance scores because of increased generalization of an area
In [5]:
factor = 1
dd = getInter(ds, factor)
print(np.unique(abs(np.diff(dd.lon.to_numpy()))),
np.unique(abs(np.diff(dd.lat.to_numpy()))))
print()
res = abs(np.diff(dd.lon.to_numpy()))[0]
display(dd)
[0.04166667 0.04166667] [0.04166667 0.04166667]
<xarray.Dataset>
Dimensions: (lon: 52, lat: 44, day: 731)
Coordinates:
* lon (lon) float64 -122.1 -122.0 -122.0 ... -120.0 -119.9
* lat (lat) float64 41.07 41.03 40.98 ... 39.36 39.32 39.28
* day (day) datetime64[ns] 2020-01-01 2020-01-02 ... 2021-12-31
Data variables:
burning_index_g (day, lat, lon) float32 ...
relative_humidity (day, lat, lon) float32 ...
air_temperature (day, lat, lon) float32 ...
wind_speed (day, lat, lon) float32 ...
Attributes:
units: Unitless
description: BI-G
long_name: bi
standard_name: bi
dimensions: lon lat time
grid_mapping: crs
coordinate_system: WGS84,EPSG:4326In [6]:
plotInter(firePer, ds)